METR Report: AI Is Now Building AI Twice as Fast — What It Means for Your Tools

Introduction: The Audit That Changed the Conversation

When Anthropic's own filings warned this week that AI might soon accelerate the building of AI, it was the company grading itself. On November 5, an outside group delivered something rarer: an independent look inside the lab. METR, the nonprofit research organization best known for measuring how long AI agents can work autonomously before needing human help, published the results of an 11-week study of Anthropic's internal development — and its headline finding is striking: Anthropic is probably improving its models about twice as fast as it did during the March 2024 to October 2025 period, and roughly 80% of that acceleration comes from AI itself, not from more computers or more hires.

For anyone choosing AI tools for their workflow, this matters more than another benchmark score. If the models powering Claude, ChatGPT and their peers are now being improved at a compounding, AI-assisted pace, the tools you buy, subscribe to, and build on will get better (and shift under you) faster than any previous software cycle. Here's what METR found, what the caveats are, and how to position your stack for it.

For background on Anthropic's own claims, see our earlier coverage: Claude Is Building Claude: Anthropic Says AI Now Leads 26% of Its Own R&D.

What METR Actually Found

METR's team examined Anthropic's internal test scores, data on computing power and staffing, staff surveys, and roughly 15 interviews over an 11-week window from August 12 through October 29, 2026. The organization was not paid for the study, and notably, it did not get to try unreleased models itself — its conclusions rest on the data Anthropic shared.

Three findings stand out:

• The pace has roughly doubled. METR estimates Anthropic's capability progress is now running at about 2x the rate of the March 2024–October 2025 period — the fastest sustained acceleration it has documented at a frontier lab.
• AI is doing most of the accelerating. About 80% of the speedup is attributable to AI tools helping researchers work faster — writing code, analyzing experiments, running evaluations — rather than to added compute or headcount.
• No hard bottleneck yet. METR found no clear obstacle that would prevent the acceleration from continuing or increasing over the next year.

Where the Speedup Comes From

The mechanism is not science fiction. Anthropic staff described what METR calls "signs of life" of AI doing large parts of research tasks on its own — drafting experiment code, triaging results, suggesting hypotheses — while humans still direct the work. AI is not yet independently conceiving and completing research projects. But the arithmetic is the story: when each researcher is effectively amplified several-fold on coding and analysis, the whole research pipeline speeds up, and the improved models those researchers ship then amplify the next round. That feedback loop is the practical definition of recursive self-improvement, arriving gradually rather than all at once.

It maps onto what's already shipping to consumers. Agentic coding tools like Claude Code, Cursor and GitHub Copilot give ordinary developers a scaled-down version of the same leverage frontier researchers use daily. The gap between "AI helps me code" and "AI helps build the next AI" is one of degree, not kind.

The Catches: Redactions and Auto-Graded Tasks

METR is careful, and its report carries two big qualifications that readers should weigh.

First, the redactions. Anthropic blacked out much of the material METR found most informative — the company says to protect confidential information that could help competitors. METR notes the public "has to rely on METR's judgment much more than would be ideal," and describes its own report as closer to a supervised audit than a full inspection. The findings back Anthropic's claims, but how firm the evidence is remains partly a matter of trust.

Second, the task mix. The measured acceleration may hold only for work that is easy to grade automatically — coding tasks, benchmarks, quantitative evaluations. Research taste, novel experimental design, and safety judgment are harder to measure, and METR's method says little about whether those are speeding up too. A doubling on auto-gradable tasks is not yet a doubling of the entire research function.

Why It Matters for the AI Tools You Use

If METR's estimate is even directionally right, three practical consequences follow for tool buyers and builders:

• Model generations will churn faster. The annual "big release" rhythm is already gone; expect meaningful capability jumps on a quarterly or faster cadence. Subscriptions that felt stable — like your ChatGPT plan — will change underneath you mid-cycle, sometimes in ways you notice.
• Open weights will trail by less. Faster frontier iteration also pulls the open ecosystem along. Open-weight families from DeepSeek, Kimi and Qwen already compress frontier capabilities onto consumer hardware within days; a faster frontier means the discount lag shrinks too.
• Evaluations age quickly. A tool comparison written three months ago may already be stale. Treat rankings — including ours — as snapshots, and re-test the shortlist before committing to an annual plan.

How to Prepare Your Tool Stack for Faster Model Cycles

You can't stop the pace, but you can stop paying for it unnecessarily:

• Prefer monthly billing over annual locks while the leaderboard is this volatile.
• Abstract your dependencies. Route through providers and gateways that let you swap underlying models without rewriting your app or prompts.
• Automate re-evaluation. Keep a small, versioned set of your own test tasks and re-run them whenever a major model updates. It's the only benchmark that reflects your workload.
• Watch the assist layer. The biggest near-term gains for most teams are still in adopting agentic coding and research workflows, not in waiting for the next frontier model.

Browse the full range of options in our AI programming tools category to see what an AI-amplified workflow looks like today.

Frequently Asked Questions